Guy G
40+ years programming experience including recent AI and ML.
New to Contra
Guy is ready for their next project!
Lots of projects in the works! Too many really.
Opus 4.5 has jammed my design episodes into hyper mode. I'm glad I heard Andrew Ng talk about the exhaustion after a day of working with agentic code helpers. That showed up months ago and I still didn't recognize.
Now, having to hold so much in my mind at once while it is also growing at a recognizable pace is noticeably stressful.
Here is my latest build, a useful agency for questions at high accuracy with 9 different agents available to work in tandem. Special thanks to my friend Enigma Specter who provided the original backend mechanisms. I salvaged his design with a nice web UI and added a web Researcher agent:
https://guyzer.pythonanywhere.com/mrce/
You will need to register but it's just an email username and password, no verification loop yet. And you'll need an API key.
1
14
Cyclical Adversarial Stepwise Improvement CASI: Live Demo
We talk frequently about AI self-correction but few projects provide a robust architectural loop for it.
CASI formalizes the process of non-linear refinement:
The Generator proposes a solution or concept.
The Critic applies structural pressure and targeted feedback.
The system cycles automatically until Structural Coherence is reached.
You can now test the CASI engine live and see how an autonomous refinement process delivers structurally superior output:
Try the CASI web app here:
https://guyzer.pythonanywhere.com/casi/
14
99
Further investigations of the Biology Latent Space.
1
9
Let's visualize some more biology. Audio by Sacrificial Pancakes and Korean tree frogs.
1
19
Each pulse consists of the changes needed to transform the visualization of one subject range to another. I could reduce the pulsing by using inter-prompt interpolation but at the time I made these last year the computational overhead was too expensive. If I had the time to do it now it would be quite a bit more efficient on an A100 for about the same cost including interpolation.
This is a visualization of a couple months worth of "Physical Review Letters E" which is very much worth checking out.
1
26
Another fun project I started last year is called "CASI" for Cyclical Adversarial Step-wise Improvement. It leverages the fact that when LLMs engage in self-correcting loops they get better at a task. In this case we are pitting 2 models and 2 system prompts with distinctly different purposes against each other to improve a concept or idea. This is unlike self-play and more like other-play. I would love to include an intermittent training pass or LoRa construction between cycles to make the models focus more completely. CASI is still a work in progress so use at your own frustration. If you encounter any issues please log them with the github repo, you will be helping the entire world :)
https://github.com/TheOneTrueGuy/CASI
22
134
Converted some Physical Review Letters B into visualizations so our collective subconscious could extract concept clusters from them visually without having to convert from word forms.
0
12
Another fun and useful project from 2024. I did this work for a nice gentleman named Andy Walters. He was looking to automatically split videos along the natural pause boundaries between sentences.
So I wrote up this neat little script that uses the Whisper-Timestamp model and library to find those boundaries and split the video on them so that no words get chopped in mid-utterance.
The company Andy was working for didn't make their next round of funding. Before we mothballed the whole thing Andy gave me permission to open source the code.
I realized too late, with just one more pass through an LLM you could have the json splitter file that was generated rearranged to split by subject. If I had the time I would include that feature but until then it is still useful like it is.
https://github.com/TheOneTrueGuy/G_video_clipper
1
17
A friend of mine does this thing on X that she calls "Whitepill Wednesday" which is a celebration of optimism and acceleration into full human flourishing. So I made this video back in February to help her and show my appreciation.
17
100
Last year, when I had a lot more time, I had a lot of fun converting biology journal entries into visualizations using stable diffusion image generation. This one turned out particularly good. I don't know if it works like I thought it might but being able to see the descriptions of biological processes seemed like an interesting way to possibly learn about biology. I did something similar with Physics Review Letters A through G and will try to get around to sharing some of those too. I even have one called "Odd Proton" that is focused on fusion processes.
17
111
A few months ago I made a software device for tinkering with LLMs.
It only works with older/smaller models and won't load gguf yet but it's got a decent visualizer and a few other items like being able to invert/modify anything from a single weight to entire layers or blocks of weights.
It was working fine a few months ago when I moved on so I thought it might be useful to other people playing with LLMs.
Hopefully it will also get a little love from the Github community to improve it further.
https://github.com/TheOneTrueGuy/0LLM_Tinkerer
15
94
Language models provide some spectacular new opportunities for discovery. While working on some ideas it occurred to me to explore the relationships between concepts spatially. Interpolating across the latent space between concepts and then mapping to a vectors nearest token predicate. I arranged this formulation as a form of tessellation to cover the n-dimensional volume efficiently. The result is a work-in-progress I call the Tessellator. Since this is an experiment it just made sense to open source it and a new update is due in the next week or two.
https://github.com/TheOneTrueGuy/tessellator
13
81
This work was from late last year but I've been meaning to revisit it with an Agentic twist. Rebuilt to remain running in real time like a guardian I think it could be truly useful. It got open-sourced in this form after the client decided not to pursue it further. It is a simple tool for analyzing enormous blocks of emails for signs of fraud or deception. It won't take much for me to elevate the performance.
https://github.com/TheOneTrueGuy/Fraud-Analysis-Tool
and here is the video that was made for the associated Kaggle contest:
https://www.youtube.com/watch?v=do9uPzW4AVk
20
118
This is a combination of Physical Review Letters B visualized by Stable Diffusion animation and a Langton's Ant automaton music audio generator.
1
17
AI-Powered Fraud Detection Platform for Kaggle Contest
0
1
I've got a new animation script in the works using stable diffusion, pure custom code. Also a nifty search laboratory that will scrape endpoints and/or data depending on your needs. My repos have a bunch of treasures to share: https://github.com/TheOneTrueGuy?tab=repositories&q=&type=source
Finally, here's an animation I made from a large collection of biology journals. A unique blend of science and artwork.
16
101